Skip to content
AI & Agents
Skill

/project-structure

Generate a single-file compressed symbol map of a TypeScript or Swift repository — files, exported symbols, typed signatures, plugin boundaries — sized to fit an LLM context window. Use for whole-project reasoning: duplication hunting, refactor planning, architecture recon, or

BOOST
From plugin
agent-scripts
7.1k54 skills
Install
$ npx -y skills add steipete/agent-scripts --skill project-structure --agent claude-code

How it fires

How this skill gets triggered: by you, by Claude, or both.

  • Fires itselfAuto-invocation. Claude auto-loads it when your prompt matches the work.Auto-invocation is when the right skill fires by itself at the right moment, driven by a FLOW.md router and a hook, instead of you invoking it by name. It is the difference between a skill being installed and a skill actually getting used.Read the full definition →
  • You can call itInvoke it directly when you want it.
  • Slash command/project-structure

Context preview

The summary Claude sees to decide when to auto-load this skill.

Generate a single-file compressed symbol map of a TypeScript or Swift repository — files, exported symbols, typed signatures, plugin boundaries — sized to fit an LLM context window. Use for whole-project reasoning: duplication hunting, refactor planning, architecture recon, or

SKILL.md

project-structure.SKILL.md
name: project-structure
description: "Generate a single-file compressed symbol map of a TypeScript or Swift repository — files, exported symbols, typed signatures, plugin boundaries — sized to fit an LLM context window. Use for whole-project reasoning: duplication hunting, refactor planning, architecture recon, or feeding another agent a full-project map."

Project Structure

Compress a TS or Swift repo into one map file an agent can load whole. Backed by `map.ts` next to this file. TS: parse-only TS compiler, no type-check; ~8k files in ~5s; resolves the `typescript` package from the target repo, falling back to this skill dir. Swift: zero-dependency regex/brace-depth scanner built into `map.ts` (no `typescript` needed for pure-Swift repos). Requires Node >= 23.6 (native type stripping) or `npx tsx`. `.ts`/`.tsx` and `.swift` files are detected by extension; a repo may mix both.

Run

node <this-skill-dir>/map.ts <repoRoot> [flags]

`<this-skill-dir>` is the base directory of this skill as announced when the skill loads (canonical: `~/Projects/agent-scripts/skills/project-structure`).

Output: one map file (default `project-structure-map.txt` in cwd) plus a JSON stats line (files, symbols, bytes, approxTokens) on stdout.

Flags

  • `--out <file>` — output path.
  • `--mode dense|skeleton|exports|sigs|full` — default `dense`.
  • `dense`: dir-grouped, one line per file: `file fn:a,b ty:T cl:C c:x re:./y`. Recon tier. File extensions are stripped for compactness, so a same-basename TS and Swift file in one dir share a line prefix (theoretical in practice; grep the repo to disambiguate).
  • `skeleton`: one symbol per line, names only.
  • `exports`: exported symbols with full typed signatures, type bodies, first doc-comment line. Refactor-decision tier. **TS-only** — Swift files fall back to their dense-style line (no fabricated signatures).
  • `sigs`: exports but type/interface bodies collapsed to member names (only ~10% smaller than exports; rarely worth it). TS-only, same Swift fallback.
  • `full`: exports + non-exported top-level symbols (marked `internal`; internal consts appear only when function-valued or explicitly typed — untyped internal consts are filtered as noise). TS-only, same Swift fallback.
  • `--include a,b,c` — paths to map, relative to repoRoot (default: all top-level dirs, minus skips/boundaries). Accepts nested paths, not just top-level dirs: `--include src/channels/turn` maps exactly that subtree. Headers stay repoRoot-relative, so running from repo root with `--include <subsystem>` is the recommended way to map a subsection.
  • `--boundary x,y` — dirs listed as boundary index only: child module names + one-line `package.json` descriptions, no symbol content. Use for plugin/extension trees.
  • `--include-tests` — keep test files and test-support infra (default: both excluded; Swift test detection is path-based: `Tests?/` dirs and `*Tests.swift`).
  • `--no-docs` — strip doc comments (saves ~13% on typed modes).
  • `--fn-consts-only` — drop non-function const exports (on Swift files: drops every let/var symbol — closure-valued consts are indistinguishable without type info). Caution: silently drops files whose only exports are consts (plugin definition objects, schemas, registries). Prefer keeping consts in `dense` (names are cheap).
  • `--re-counts` — collapse re-export path lists to a count (dense; saves ~10%).
  • `--max-per-kind N` — cap each symbol-kind list per file with `+n` overflow marker (dense; 10–12 is safe).
  • `--members` — Swift only (ignored for TS files): additionally emit one-level-deep methods/properties as `Type.member`; extension members appear under their target type.
  • `--public-only` — Swift only (ignored for TS files): keep only symbols whose written access is `open`/`public`.

When `--include` is absent, source files sitting directly in repoRoot are scanned too (so pointing repoRoot at a leaf source dir works); with `--include`, only the listed paths are walked.

Swift specifics

Zero-dependency line scanner (lexical masking of comments/strings + brace-depth tracking), validated 1:1 against SourceKitten on 1,850 files across two real repos (100% top-level symbol agreement) at ~100× SourceKitten's speed. Known limitations (accepted tradeoffs of the parserless design): one declaration per line is assumed — a member on the same line as its type (`struct Box { var v = 1 }`) or a second semicolon-separated declaration is not emitted; bare `/…/` regex literals are not masked (lexically ambiguous with division), so a brace inside one can desync depth for the rest of that file — extended `#/…/#` literals are masked correctly; comments or raw/multiline literals nested *inside* string interpolations, emoji identifiers, and `#if` branches whose alternative headers declare *differently named* containers are best-effort (same-name platform-split containers work). Kinds: `fn` = func (incl. operator funcs); `cl` = class, actor; `ty` = struct, enum, protocol, typealias, macro, and extensions as `extension:TargetType`; `c` = let/var. Visibility is appended per symbol as `[open]`/`[public]`/`[package]`/`[private]`/`[fileprivate]`; `internal` (Swift's default) is deliberately untagged to save map bytes. Tags reflect *written* access; members of protocols/extensions without a written modifier inherit the container's access. Swift puts most code inside types, so the default top-level map is thin (median ~2 symbols/file) — reach for `--members` when method-level recon matters.

Sizing

TS (reference: openclaw, ~7M LOC, ~14k source files; o200k tokens; byte/4 estimate runs ~5–15% high — verify with a real tokenizer when near a budget):

  • exports, whole repo: ~2.5M tokens — never fits; scope typed maps to one subsystem.
  • exports, one subsystem (e.g. src/channels, 257 files): ~79k.
  • skeleton, whole repo: ~717k (fits 1M-class windows).
  • dense, whole repo: ~430k real.
  • dense, `src`+`packages` + extensions boundary, `--re-counts --max-per-ki
Read more
Ships withagent-scripts

Shared agent instructions, skills, and small portable helpers for Peter's local workspaces.

Get the whole plugin
Stats
7,208
Stars
617
Forks
Active
Maintenance
Shell
Language
MIT
License
19h ago
Last commit
10mo ago
Created
3h ago
Added

Repo: steipete/agent-scripts

Other skills on agent-scripts.